Economically motivated meta-learning for multi-step stock return forecasting: evidence from global equity markets

Forecasting stock returns over multiple horizons is an important issue in financial econometrics, but existing trajectory-based meta-learning methods pay limited attention to economically interpretable features of forecast paths. This paper proposes Forecasted Trajectory Economically-aware Neighbours (FTEN), an economically motivated refinement of the Forecasted Trajectory Neighbours framework that selects historical neighbours based on similarity in both the average-return level and the demeaned residual shape of forecast paths. Using daily returns from 16 major national equity markets, we show that FTEN delivers more accurate forecasts and higher investor utility than existing trajectory-based approaches, with robust gains across markets, horizons, and alternative specifications. These findings highlight that economically motivated restrictions on trajectory matching can improve both predictive performance and economic relevance in multi-step return forecasting.